FakeQuantWithMinMaxVarsPerChannelGradient

tensorflow C++ API

tensorflow::ops::FakeQuantWithMinMaxVarsPerChannelGradient

Compute gradients for a FakeQuantWithMinMaxVarsPerChannel operation.

Summary

Arguments:

• scope: A Scope object
• gradients: Backpropagated gradients above the FakeQuantWithMinMaxVars operation, shape one of: [d], [b, d], [b, h, w, d] .
• inputs: Values passed as inputs to the FakeQuantWithMinMaxVars operation, shape same as gradients Quantization interval, floats of shape [d] .

Optional attributes (seeAttrs):

• num_bits: The bitwidth of the quantization; between 2 and 8, inclusive.
• narrow_range: Whether to quantize into 2^num_bits - 1 distinct values.

Returns:

• Outputbackprops_wrt_input: Backpropagated gradients w.r.t. inputs, shape same as inputs : gradients * (inputs >= min && inputs <= max).
• Output backprop_wrt_min: Backpropagated gradients w.r.t. min parameter, shape [d] : sum_per_d(gradients * (inputs < min)) .
• Output backprop_wrt_max: Backpropagated gradients w.r.t. max parameter, shape [d] : sum_per_d(gradients * (inputs > max)).

FakeQuantWithMinMaxVarsPerChannelGradient block

Source link :https://github.com/EXPNUNI/enuSpaceTensorflow/blob/master/enuSpaceTensorflow/tf_array_ops.cpp

Argument:

• Scope scope : A Scope object (A scope is generated automatically each page. A scope is not connected.)
• Input inputs: A Tensor of type float.
• Input min : A Tensor of type float.
• Input max : A Tensor of type float.
• Attr attrs : An optional attribute value
  • num_bits : An optional int. Defaults to 8.

Attrs use ex)

Output:

• output : Output object of FakeQuantWithMinMaxVarsPerChannelGradient class object.

Result:

• std::vector(Tensor) result_output: A Tensor of type float. This operation has a gradient and thus allows for training min and max values.

Using Method